Mira Murati's 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis' AI FINRA | EP #271 — Key Takeaways
YouTube
Mira Murati's 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis' AI FINRA | EP #271
Peter H. Diamandis1h 57mJul 17, 2026
Watch the originalLiquid AI's Ramine Hassani made the strongest case that true recursive self-improvement requires weight-level neural adaptation — not prompt/code engineering loops — and that current "breakthroughs" like Weco AI's system would take 350 years to fine-tune a 2-billion-parameter model under chinchilla scaling laws.
Key takeaways
Liquid AI's Mercedes model is under 1GB, runs on a $60 chip with no cloud connection
Liquid AI's Mercedes model is under 1GB, runs on a $60 chip with no cloud connection
- Deployed across all Mercedes-Benz North America cars from 2022 onward via a 600MB OTA update this year
- Controls 700–1,200 car functions; updates can be as small as 20MB via LoRA-style adapters
Ramine: WICO's 'recursive self-improvement' doesn't change model weights — it's prompt engineering
Ramine: WICO's 'recursive self-improvement' doesn't change model weights — it's prompt engineering
- No neural network weight updates occur in the pipeline; capabilities of underlying models remain fixed
- Using Chinchilla scaling laws, their framework would take 350 years to fine-tune a 2B parameter model
West lags on openweight models because closed APIs are too profitable to abandon
West lags on openweight models because closed APIs are too profitable to abandon
- Anthropic and OpenAI each targeting ~$1T IPO valuations on per-token API revenue — no incentive to release weights
- China, GPU-deprived and CCP-mandated to integrate AI societally, monetizes via applications and chips instead
This Dig holds 4 more insights, 4 flashcards, and 3 quotes — free in Homestake.
Unlock this Dig freeFree forever · No credit card required
In this video
- 1mEpisode Intro: Murati, Liquid AI, and Regulation
- 4mAI regulation and standards bodies
- 20mUS-China AI capability framework
- 26mBreakthrough in open weight models with Thinking Machine Labs
- 43mRecursive self-improvement and AI safety
- 49mThe future of AI model development and regulation
- 55mWomen in AI leadership and industry diversity
- 58mEmerging Capabilities of Foundation Models
- 1h 20mSmall Language Models and On-Device AI
- 1h 27mAI in Automotive: Mercedes Partnership
- 1h 38mArchitectures Beyond Transformers
- 1h 53mAI and Longevity: Reversing Aging with Enzymes
- 1h 56mThe Future of Human and Organizational Evolution
“Customization over leaderboard dominance uh is what's going to win her the day.”
This page is a partial, transformative summary produced by Homestake. All rights to the original content remain with its creator — please support them at the source link above.
